# Copyright 2011-2016 The OpenSSL Project Authors. All Rights Reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Generated by admin/import-aws-lc from aws-lc 91c93a8eb860e51482a871a8772cb9b1b310baea:
# generated-src/linux-x86_64/crypto/fipsmodule/x86_64-mont5.S, which aws-lc generates from crypto/fipsmodule/bn/asm/x86_64-mont5.pl.
# Do not edit by hand.
.align 32
.cfi_startproc
movq %rsp,%rax
.cfi_def_cfa_register %rax
pushq %rbx
.cfi_offset %rbx,-16
pushq %rbp
.cfi_offset %rbp,-24
pushq %r12
.cfi_offset %r12,-32
pushq %r13
.cfi_offset %r13,-40
pushq %r14
.cfi_offset %r14,-48
pushq %r15
.cfi_offset %r15,-56
.Lmont5_power5_prologue:
shll $3,%r9d
leal (%r9,%r9,2),%r10d
negq %r9
movq (%r8),%r8
leaq -320(%rsp,%r9,2),%r11
movq %rsp,%rbp
subq %rdi,%r11
andq $4095,%r11
cmpq %r11,%r10
jb .Lmont5_pwr_sp_alt
subq %r11,%rbp
leaq -320(%rbp,%r9,2),%rbp
jmp .Lmont5_pwr_sp_done
.align 32
.Lmont5_pwr_sp_alt:
leaq 4096-320(,%r9,2),%r10
leaq -320(%rbp,%r9,2),%rbp
subq %r10,%r11
movq $0,%r10
cmovcq %r10,%r11
subq %r11,%rbp
.Lmont5_pwr_sp_done:
andq $-64,%rbp
movq %rsp,%r11
subq %rbp,%r11
andq $-4096,%r11
leaq (%r11,%rbp,1),%rsp
movq (%rsp),%r10
cmpq %rbp,%rsp
ja .Lmont5_pwr_page_walk
jmp .Lmont5_pwr_page_walk_done
.Lmont5_pwr_page_walk:
leaq -4096(%rsp),%rsp
movq (%rsp),%r10
cmpq %rbp,%rsp
ja .Lmont5_pwr_page_walk
.Lmont5_pwr_page_walk_done:
movq %r9,%r10
negq %r9
movq %r8,32(%rsp)
movq %rax,40(%rsp)
.cfi_escape 0x0f,0x05,0x77,0x28,0x06,0x23,0x08
.Lmont5_power5_body:
.byte 102,72,15,110,207
.byte 102,72,15,110,209
.byte 102,73,15,110,218
.byte 102,72,15,110,226
call {bn_sqr8x_internal}
call .Lmont5___bn_post4x_internal
call {bn_sqr8x_internal}
call .Lmont5___bn_post4x_internal
call {bn_sqr8x_internal}
call .Lmont5___bn_post4x_internal
call {bn_sqr8x_internal}
call .Lmont5___bn_post4x_internal
call {bn_sqr8x_internal}
call .Lmont5___bn_post4x_internal
.byte 102,72,15,126,209
.byte 102,72,15,126,226
movq %rsi,%rdi
movq 40(%rsp),%rax
leaq 32(%rsp),%r8
call .Lmont5_mul4x_internal
movq 40(%rsp),%rsi
.cfi_def_cfa %rsi,8
movq $1,%rax
movq -48(%rsi),%r15
.cfi_restore %r15
movq -40(%rsi),%r14
.cfi_restore %r14
movq -32(%rsi),%r13
.cfi_restore %r13
movq -24(%rsi),%r12
.cfi_restore %r12
movq -16(%rsi),%rbp
.cfi_restore %rbp
movq -8(%rsi),%rbx
.cfi_restore %rbx
leaq (%rsi),%rsp
.cfi_def_cfa_register %rsp
.Lmont5_power5_epilogue:
.byte 0xf3,0xc3
.cfi_endproc
.align 32
.Lmont5_mul4x_internal:
.cfi_startproc
shlq $5,%r9
movd 8(%rax),%xmm5
leaq {inc}(%rip),%rax
leaq 128(%rdx,%r9,1),%r13
shrq $5,%r9
movdqa 0(%rax),%xmm0
movdqa 16(%rax),%xmm1
leaq 88-112(%rsp,%r9,1),%r10
leaq 128(%rdx),%r12
pshufd $0,%xmm5,%xmm5
movdqa %xmm1,%xmm4
.byte 0x67,0x67
movdqa %xmm1,%xmm2
paddd %xmm0,%xmm1
pcmpeqd %xmm5,%xmm0
.byte 0x67
movdqa %xmm4,%xmm3
paddd %xmm1,%xmm2
pcmpeqd %xmm5,%xmm1
movdqa %xmm0,112(%r10)
movdqa %xmm4,%xmm0
paddd %xmm2,%xmm3
pcmpeqd %xmm5,%xmm2
movdqa %xmm1,128(%r10)
movdqa %xmm4,%xmm1
paddd %xmm3,%xmm0
pcmpeqd %xmm5,%xmm3
movdqa %xmm2,144(%r10)
movdqa %xmm4,%xmm2
paddd %xmm0,%xmm1
pcmpeqd %xmm5,%xmm0
movdqa %xmm3,160(%r10)
movdqa %xmm4,%xmm3
paddd %xmm1,%xmm2
pcmpeqd %xmm5,%xmm1
movdqa %xmm0,176(%r10)
movdqa %xmm4,%xmm0
paddd %xmm2,%xmm3
pcmpeqd %xmm5,%xmm2
movdqa %xmm1,192(%r10)
movdqa %xmm4,%xmm1
paddd %xmm3,%xmm0
pcmpeqd %xmm5,%xmm3
movdqa %xmm2,208(%r10)
movdqa %xmm4,%xmm2
paddd %xmm0,%xmm1
pcmpeqd %xmm5,%xmm0
movdqa %xmm3,224(%r10)
movdqa %xmm4,%xmm3
paddd %xmm1,%xmm2
pcmpeqd %xmm5,%xmm1
movdqa %xmm0,240(%r10)
movdqa %xmm4,%xmm0
paddd %xmm2,%xmm3
pcmpeqd %xmm5,%xmm2
movdqa %xmm1,256(%r10)
movdqa %xmm4,%xmm1
paddd %xmm3,%xmm0
pcmpeqd %xmm5,%xmm3
movdqa %xmm2,272(%r10)
movdqa %xmm4,%xmm2
paddd %xmm0,%xmm1
pcmpeqd %xmm5,%xmm0
movdqa %xmm3,288(%r10)
movdqa %xmm4,%xmm3
paddd %xmm1,%xmm2
pcmpeqd %xmm5,%xmm1
movdqa %xmm0,304(%r10)
paddd %xmm2,%xmm3
.byte 0x67
pcmpeqd %xmm5,%xmm2
movdqa %xmm1,320(%r10)
pcmpeqd %xmm5,%xmm3
movdqa %xmm2,336(%r10)
pand 64(%r12),%xmm0
pand 80(%r12),%xmm1
pand 96(%r12),%xmm2
movdqa %xmm3,352(%r10)
pand 112(%r12),%xmm3
por %xmm2,%xmm0
por %xmm3,%xmm1
movdqa -128(%r12),%xmm4
movdqa -112(%r12),%xmm5
movdqa -96(%r12),%xmm2
pand 112(%r10),%xmm4
movdqa -80(%r12),%xmm3
pand 128(%r10),%xmm5
por %xmm4,%xmm0
pand 144(%r10),%xmm2
por %xmm5,%xmm1
pand 160(%r10),%xmm3
por %xmm2,%xmm0
por %xmm3,%xmm1
movdqa -64(%r12),%xmm4
movdqa -48(%r12),%xmm5
movdqa -32(%r12),%xmm2
pand 176(%r10),%xmm4
movdqa -16(%r12),%xmm3
pand 192(%r10),%xmm5
por %xmm4,%xmm0
pand 208(%r10),%xmm2
por %xmm5,%xmm1
pand 224(%r10),%xmm3
por %xmm2,%xmm0
por %xmm3,%xmm1
movdqa 0(%r12),%xmm4
movdqa 16(%r12),%xmm5
movdqa 32(%r12),%xmm2
pand 240(%r10),%xmm4
movdqa 48(%r12),%xmm3
pand 256(%r10),%xmm5
por %xmm4,%xmm0
pand 272(%r10),%xmm2
por %xmm5,%xmm1
pand 288(%r10),%xmm3
por %xmm2,%xmm0
por %xmm3,%xmm1
por %xmm1,%xmm0
pshufd $0x4e,%xmm0,%xmm1
por %xmm1,%xmm0
leaq 256(%r12),%r12
.byte 102,72,15,126,195
movq %r13,16+8(%rsp)
movq %rdi,56+8(%rsp)
movq (%r8),%r8
movq (%rsi),%rax
leaq (%rsi,%r9,1),%rsi
negq %r9
movq %r8,%rbp
mulq %rbx
movq %rax,%r10
movq (%rcx),%rax
imulq %r10,%rbp
leaq 64+8(%rsp),%r14
movq %rdx,%r11
mulq %rbp
addq %rax,%r10
movq 8(%rsi,%r9,1),%rax
adcq $0,%rdx
movq %rdx,%rdi
mulq %rbx
addq %rax,%r11
movq 8(%rcx),%rax
adcq $0,%rdx
movq %rdx,%r10
mulq %rbp
addq %rax,%rdi
movq 16(%rsi,%r9,1),%rax
adcq $0,%rdx
addq %r11,%rdi
leaq 32(%r9),%r15
leaq 32(%rcx),%rcx
adcq $0,%rdx
movq %rdi,(%r14)
movq %rdx,%r13
jmp .Lmont5_1st4x
.align 32
.Lmont5_1st4x:
mulq %rbx
addq %rax,%r10
movq -16(%rcx),%rax
leaq 32(%r14),%r14
adcq $0,%rdx
movq %rdx,%r11
mulq %rbp
addq %rax,%r13
movq -8(%rsi,%r15,1),%rax
adcq $0,%rdx
addq %r10,%r13
adcq $0,%rdx
movq %r13,-24(%r14)
movq %rdx,%rdi
mulq %rbx
addq %rax,%r11
movq -8(%rcx),%rax
adcq $0,%rdx
movq %rdx,%r10
mulq %rbp
addq %rax,%rdi
movq (%rsi,%r15,1),%rax
adcq $0,%rdx
addq %r11,%rdi
adcq $0,%rdx
movq %rdi,-16(%r14)
movq %rdx,%r13
mulq %rbx
addq %rax,%r10
movq 0(%rcx),%rax
adcq $0,%rdx
movq %rdx,%r11
mulq %rbp
addq %rax,%r13
movq 8(%rsi,%r15,1),%rax
adcq $0,%rdx
addq %r10,%r13
adcq $0,%rdx
movq %r13,-8(%r14)
movq %rdx,%rdi
mulq %rbx
addq %rax,%r11
movq 8(%rcx),%rax
adcq $0,%rdx
movq %rdx,%r10
mulq %rbp
addq %rax,%rdi
movq 16(%rsi,%r15,1),%rax
adcq $0,%rdx
addq %r11,%rdi
leaq 32(%rcx),%rcx
adcq $0,%rdx
movq %rdi,(%r14)
movq %rdx,%r13
addq $32,%r15
jnz .Lmont5_1st4x
mulq %rbx
addq %rax,%r10
movq -16(%rcx),%rax
leaq 32(%r14),%r14
adcq $0,%rdx
movq %rdx,%r11
mulq %rbp
addq %rax,%r13
movq -8(%rsi),%rax
adcq $0,%rdx
addq %r10,%r13
adcq $0,%rdx
movq %r13,-24(%r14)
movq %rdx,%rdi
mulq %rbx
addq %rax,%r11
movq -8(%rcx),%rax
adcq $0,%rdx
movq %rdx,%r10
mulq %rbp
addq %rax,%rdi
movq (%rsi,%r9,1),%rax
adcq $0,%rdx
addq %r11,%rdi
adcq $0,%rdx
movq %rdi,-16(%r14)
movq %rdx,%r13
leaq (%rcx,%r9,1),%rcx
xorq %rdi,%rdi
addq %r10,%r13
adcq $0,%rdi
movq %r13,-8(%r14)
jmp .Lmont5_outer4x
.align 32
.Lmont5_outer4x:
leaq 16+128(%r14),%rdx
pxor %xmm4,%xmm4
pxor %xmm5,%xmm5
movdqa -128(%r12),%xmm0
movdqa -112(%r12),%xmm1
movdqa -96(%r12),%xmm2
movdqa -80(%r12),%xmm3
pand -128(%rdx),%xmm0
pand -112(%rdx),%xmm1
por %xmm0,%xmm4
pand -96(%rdx),%xmm2
por %xmm1,%xmm5
pand -80(%rdx),%xmm3
por %xmm2,%xmm4
por %xmm3,%xmm5
movdqa -64(%r12),%xmm0
movdqa -48(%r12),%xmm1
movdqa -32(%r12),%xmm2
movdqa -16(%r12),%xmm3
pand -64(%rdx),%xmm0
pand -48(%rdx),%xmm1
por %xmm0,%xmm4
pand -32(%rdx),%xmm2
por %xmm1,%xmm5
pand -16(%rdx),%xmm3
por %xmm2,%xmm4
por %xmm3,%xmm5
movdqa 0(%r12),%xmm0
movdqa 16(%r12),%xmm1
movdqa 32(%r12),%xmm2
movdqa 48(%r12),%xmm3
pand 0(%rdx),%xmm0
pand 16(%rdx),%xmm1
por %xmm0,%xmm4
pand 32(%rdx),%xmm2
por %xmm1,%xmm5
pand 48(%rdx),%xmm3
por %xmm2,%xmm4
por %xmm3,%xmm5
movdqa 64(%r12),%xmm0
movdqa 80(%r12),%xmm1
movdqa 96(%r12),%xmm2
movdqa 112(%r12),%xmm3
pand 64(%rdx),%xmm0
pand 80(%rdx),%xmm1
por %xmm0,%xmm4
pand 96(%rdx),%xmm2
por %xmm1,%xmm5
pand 112(%rdx),%xmm3
por %xmm2,%xmm4
por %xmm3,%xmm5
por %xmm5,%xmm4
pshufd $0x4e,%xmm4,%xmm0
por %xmm4,%xmm0
leaq 256(%r12),%r12
.byte 102,72,15,126,195
movq (%r14,%r9,1),%r10
movq %r8,%rbp
mulq %rbx
addq %rax,%r10
movq (%rcx),%rax
adcq $0,%rdx
imulq %r10,%rbp
movq %rdx,%r11
movq %rdi,(%r14)
leaq (%r14,%r9,1),%r14
mulq %rbp
addq %rax,%r10
movq 8(%rsi,%r9,1),%rax
adcq $0,%rdx
movq %rdx,%rdi
mulq %rbx
addq %rax,%r11
movq 8(%rcx),%rax
adcq $0,%rdx
addq 8(%r14),%r11
adcq $0,%rdx
movq %rdx,%r10
mulq %rbp
addq %rax,%rdi
movq 16(%rsi,%r9,1),%rax
adcq $0,%rdx
addq %r11,%rdi
leaq 32(%r9),%r15
leaq 32(%rcx),%rcx
adcq $0,%rdx
movq %rdx,%r13
jmp .Lmont5_inner4x
.align 32
.Lmont5_inner4x:
mulq %rbx
addq %rax,%r10
movq -16(%rcx),%rax
adcq $0,%rdx
addq 16(%r14),%r10
leaq 32(%r14),%r14
adcq $0,%rdx
movq %rdx,%r11
mulq %rbp
addq %rax,%r13
movq -8(%rsi,%r15,1),%rax
adcq $0,%rdx
addq %r10,%r13
adcq $0,%rdx
movq %rdi,-32(%r14)
movq %rdx,%rdi
mulq %rbx
addq %rax,%r11
movq -8(%rcx),%rax
adcq $0,%rdx
addq -8(%r14),%r11
adcq $0,%rdx
movq %rdx,%r10
mulq %rbp
addq %rax,%rdi
movq (%rsi,%r15,1),%rax
adcq $0,%rdx
addq %r11,%rdi
adcq $0,%rdx
movq %r13,-24(%r14)
movq %rdx,%r13
mulq %rbx
addq %rax,%r10
movq 0(%rcx),%rax
adcq $0,%rdx
addq (%r14),%r10
adcq $0,%rdx
movq %rdx,%r11
mulq %rbp
addq %rax,%r13
movq 8(%rsi,%r15,1),%rax
adcq $0,%rdx
addq %r10,%r13
adcq $0,%rdx
movq %rdi,-16(%r14)
movq %rdx,%rdi
mulq %rbx
addq %rax,%r11
movq 8(%rcx),%rax
adcq $0,%rdx
addq 8(%r14),%r11
adcq $0,%rdx
movq %rdx,%r10
mulq %rbp
addq %rax,%rdi
movq 16(%rsi,%r15,1),%rax
adcq $0,%rdx
addq %r11,%rdi
leaq 32(%rcx),%rcx
adcq $0,%rdx
movq %r13,-8(%r14)
movq %rdx,%r13
addq $32,%r15
jnz .Lmont5_inner4x
mulq %rbx
addq %rax,%r10
movq -16(%rcx),%rax
adcq $0,%rdx
addq 16(%r14),%r10
leaq 32(%r14),%r14
adcq $0,%rdx
movq %rdx,%r11
mulq %rbp
addq %rax,%r13
movq -8(%rsi),%rax
adcq $0,%rdx
addq %r10,%r13
adcq $0,%rdx
movq %rdi,-32(%r14)
movq %rdx,%rdi
mulq %rbx
addq %rax,%r11
movq %rbp,%rax
movq -8(%rcx),%rbp
adcq $0,%rdx
addq -8(%r14),%r11
adcq $0,%rdx
movq %rdx,%r10
mulq %rbp
addq %rax,%rdi
movq (%rsi,%r9,1),%rax
adcq $0,%rdx
addq %r11,%rdi
adcq $0,%rdx
movq %r13,-24(%r14)
movq %rdx,%r13
movq %rdi,-16(%r14)
leaq (%rcx,%r9,1),%rcx
xorq %rdi,%rdi
addq %r10,%r13
adcq $0,%rdi
addq (%r14),%r13
adcq $0,%rdi
movq %r13,-8(%r14)
cmpq 16+8(%rsp),%r12
jb .Lmont5_outer4x
xorq %rax,%rax
subq %r13,%rbp
adcq %r15,%r15
orq %r15,%rdi
subq %rdi,%rax
leaq (%r14,%r9,1),%rbx
movq (%rcx),%r12
leaq (%rcx),%rbp
movq %r9,%rcx
sarq $3+2,%rcx
movq 56+8(%rsp),%rdi
decq %r12
xorq %r10,%r10
movq 8(%rbp),%r13
movq 16(%rbp),%r14
movq 24(%rbp),%r15
jmp .Lmont5_sqr4x_sub_entry
.cfi_endproc
.align 32
.Lmont5___bn_post4x_internal:
.cfi_startproc
movq 0(%rbp),%r12
leaq (%rdi,%r9,1),%rbx
movq %r9,%rcx
.byte 102,72,15,126,207
negq %rax
.byte 102,72,15,126,206
sarq $3+2,%rcx
decq %r12
xorq %r10,%r10
movq 8(%rbp),%r13
movq 16(%rbp),%r14
movq 24(%rbp),%r15
jmp .Lmont5_sqr4x_sub_entry
.align 16
.Lmont5_sqr4x_sub:
movq 0(%rbp),%r12
movq 8(%rbp),%r13
movq 16(%rbp),%r14
movq 24(%rbp),%r15
.Lmont5_sqr4x_sub_entry:
leaq 32(%rbp),%rbp
notq %r12
notq %r13
notq %r14
notq %r15
andq %rax,%r12
andq %rax,%r13
andq %rax,%r14
andq %rax,%r15
negq %r10
adcq 0(%rbx),%r12
adcq 8(%rbx),%r13
adcq 16(%rbx),%r14
adcq 24(%rbx),%r15
movq %r12,0(%rdi)
leaq 32(%rbx),%rbx
movq %r13,8(%rdi)
sbbq %r10,%r10
movq %r14,16(%rdi)
movq %r15,24(%rdi)
leaq 32(%rdi),%rdi
incq %rcx
jnz .Lmont5_sqr4x_sub
movq %r9,%r10
negq %r9
.byte 0xf3,0xc3
.cfi_endproc